OBSL-1019/1007 (soda-core part): raw feature-config passthroughs on the dataset-config DTO + MM/dialect seams

[mivds]

PR chain (soda-core)

1. OBSL-1013: ship table discovery in soda-core (soda data-source discover) #2773 — ship table/column discovery
2. OBSL-1025: emit metrics[] in scan-results payload when measurements present #2774 — emit metrics[] in scan-results payload
3. OBSL-1027: SQL AST window functions + UNION-in-CTE #2776 — SQL AST window functions + UNION-in-CTE
4. OBSL-1019/1007 (soda-core part): raw feature-config passthroughs on the dataset-config DTO + MM/dialect seams #2777 — raw feature-config passthrough (config DTO) ← this PR
5. All-datasource dialect groundwork: time/interval, percentile, typed timestamp literal, NaN filter #2787 — all-datasource dialect groundwork

---

Summary

Cloud/dialect seams for the observability extensions (profiling, metric monitoring), stacked on #2776 (OBSL-1027).

Dataset-config DTO: raw feature-config passthroughs

• DatasetConfigurationDTO.profiling_configuration (alias profilingConfiguration) is a raw Optional[dict] passthrough: the soda-profiling extension (soda_profiling.profiling_config, soda-extensions) owns the parsing, so soda-core stays free of profiling semantics — the same convention as metricMonitoringConfiguration (metric-monitoring's monitor_config) and timePartitionConfiguration (the partition detector). Wire JSON unchanged; the dict arrives and round-trips verbatim. The earlier typed ProfilingConfigurationDTO + sampling-strategy union from this branch was removed again in favor of the passthrough; the typed models now live in soda-extensions.
• metric_monitoring_configuration raw list[dict] field + historic-data queries + identity helpers (MM transport seams, OBSL-1007) and get_last_measurement single-identity historic query (OBSL-1028).
• Per-CheckResult cloud-dict override hook for non-contract check shapes (OBSL-1008).

SQL AST + dialect seams

• TIME_DELTA + ADD_INTERVAL time-bucket nodes with snowflake/bigquery renderers (OBSL-1028).
• PERCENTILE_WITHIN_GROUP ordered-set aggregate node + bigquery APPROX_QUANTILES renderer (OBSL-1022).
• sql_expr_timestamp_coerce dialect hook for partition-window comparisons and get_large_numeric_cast_type_name capability for high-precision aggregate math (OBSL-1005).
• Fixes: BigQuery renders UNION as UNION ALL; BigQuery DATETIME/BIGNUMERIC + parameterized types in the soda-type reverse map; defensive Snowflake string-type aliases (OBSL-1005).

soda-core part of OBSL-1019 and the MM/dialect seams it stacks with. The profiling orchestrator + soda profile CLI (and the typed profiling-config models) are the soda-extensions part (sodadata/soda-extensions#457).

Test Plan

• Unit: profilingConfiguration passes through as a raw dict via alias, round-trips by alias unchanged, omitted key parses to None (no regression); typed parsing is covered in soda-extensions (soda-profiling/tests/test_profiling_config.py)
• pytest soda-tests/tests/unit -q: 1059 passed, 1 skipped; pre-commit clean

🤖 Generated with Claude Code

[m1n0]

Reviewed (aware it's still draft) as a mixed PR, per group. Most of it is clean: DTO passthroughs round-trip by alias and the ownership rationale is documented; observability_identity digests are pinned and the add-order is called out as load-bearing; the historic-query trio (XOR validation, 500-identity batching, grouping) is well-tested; and both override hooks (build_soda_cloud_check_dict, summary-table extra columns/headers) provably leave the default contract paths unchanged. CI green.

One finding I'd fix before undrafting:

• BigQuery build_union_sql renders plain UNION as UNION ALL (bigquery_data_source.py:174). The docstring's justification is profiling-specific ("disjoint constant metric_ labels"), but this is a general dialect method: any future caller building UNION(...) on BigQuery expecting set semantics silently gets duplicates. BigQuery supports UNION DISTINCT — rendering UNION → UNION DISTINCT and UNION_ALL → UNION ALL is both valid BQ SQL and semantics-preserving, and callers that want the cheaper form can say UNION_ALL explicitly.

Smaller notes:

• get_historic_measurements/get_historic_check_results with limit + >500 identities applies the limit per batch — merged results can exceed it. Current callers won't hit it (limit is used with single identities), but worth a line in the docstring or a guard.
• Snowflake _build_time_delta_sql counts crossed second-boundaries (TIMESTAMPDIFF(second, ...)) while the base form measures true duration — sub-second window bounds can bucket off-by-one vs postgres. Immaterial if MM scan times are whole-second aligned; flagging the boundary.
• snowflake_data_source.py:185: the new get_large_numeric_cast_type_name landed between the pre-existing "TODO: test this thorough / AI-generated" comment and the type-map function it describes — the TODO now reads as if it's about the new method.
• TIME_BUCKET_UNITS and BigQuery's _TIME_BUCKET_UNIT_NAMES are two sources of truth; adding a unit to the former silently KeyErrors at BQ render time.

[mivds]

Addressed in 23a9ae9:

• BigQuery build_union_sql — plain UNION now renders UNION DISTINCT (was UNION ALL), preserving set semantics for any caller; UNION_ALL still renders UNION ALL. Tests updated to pin both.
• Snowflake TODO placement — moved the "AI-generated / test thorough" TODO back above the type-map function it describes (the new get_large_numeric_cast_type_name had been inserted between them).
• Per-batch limit — documented on both get_historic_measurements/get_historic_check_results that limit is applied per identity-batch, so a multi-batch merged result can exceed it (current callers pass a limit only with single identities).

Deferred (noted, not changed): the Snowflake _build_time_delta_sql second-boundary note is immaterial for whole-second MM scan times; the TIME_BUCKET_UNITS vs _TIME_BUCKET_UNIT_NAMES two-sources-of-truth is the same cross-dialect consolidation you flagged on #2787 — tracking it there as one follow-up rather than touching ~6 dialects in this PR.

[sonarqubecloud]

Quality Gate passed

Issues
8 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud